A few days ago my partner Pablinux told them how to locally install the trendy Artificial Intelligence model. In this post I will list what I consider to be better models than DeepSeek and how to install and run them on our computer.
Leaving aside any political sympathies or antipathies, the Chinese government's move was a masterpiece of diplomatic marketing worthy of Sun Tzu. In contrast to Donald Trump's "Elephant in the china shop" style, they announced a model that offers the same features as ChatGPT for free and consumes fewer resources. Only those of us who follow the topic know that There have been many other open source models (some from North American companies such as Meta) for a long time, and DeepSeek's performance is only comparable to ChatGPT in the most common 5% of uses.
Large-scale language models
ChatGPT, DeepSeek and others are called Large Scale Language Models. Basically They allow a user to interact with a computer in a language similar to that used to communicate with another human being. To achieve this, they are trained with large amounts of text and rules that allow them to produce new information from what they already have.
Its main use is to answer questions, summarize texts, make translations and reproduce content.
Better models than DeepSeek and how to install them locally
Like Pablinux, we are going to use Ollama. This is a tool that allows us to install, uninstall and use different open source models from the Linux terminal. In some cases the browser can be used as a graphical interface, but we will not cover that in this article.
For Ollama to provide a proper user experience, it is best to have a dedicated GPU.Especially on the models with more parameters. However, the less powerful ones can be used on a Raspberry Pi and when I even tested models with 7 billion parameters on a computer with 6 gigs and no dedicated GPU, the computer ran without any hiccups. The same did not happen with a 13 billion one.
Parameters are the rules that the model uses to build relationships and construct patterns among the data. The more parameters and data, the more powerful a model will be; those with fewer parameters speak Spanish like Tarzan.
We can install Ollama with the commands
sudo apt install curl
curl -fsSL https://ollama.com/install.sh | sh
We can install the model with the command:
ollama pull nombre_del modelo
And run it with:
ollama run nombre_del_modelo
We uninstall it using:
ollama rm nombre_del_modelo
We can see the installed models by typing:
ollama list
The following is a small list of the models that I find most interesting: The complete list of available models can be found here here:
llama2-uncensored
Llama is a general purpose model created by Meta. In this version all restrictions introduced by the developers of the original project for legal or political reasons were removed.It has two versions, a light one that manages with 8GB and the full one that needs 64. It can be used to answer questions, write texts or in coding tasks.
Installs with:
ollama pull llama2-uncensored
And it runs with:
ollama run llama2-uncensored
codegemma
CodeGemma is a selection of lightweight yet powerful templates that allow you to perform a variety of programming tasks how to complete code or write it from scratch. Understands natural language, can follow instructions, and do mathematical reasoning.
It comes in 3 variants:
- Instruct: It transforms natural language into code and can follow instructions:
- Code: Complete and generate code from parts of existing code.
- 2b: Faster code completion task.
Tynillama
As its name indicates, it is a smaller version of the original Meta model.So it won't have as good results, but if you want to see how an AI model works on modest hardware, it's worth trying. It only has 1100 billion parameters.
Using models locally has the advantages of privacy and access to uncensored and unbiased versions that in some cases tend to be ridiculous. Microsoft's AI refused to create the image of a dachshund for me because it considered the term "bitch" offensive. The biggest drawback is the hardware requirements. It will be a matter of testing the models and finding one that is good enough for what you need and can run on the computer you have.